Add an option "cgroup_nodelete" to not delete cgroups after job completion
[fio.git] / log.c
... / ...
CommitLineData
1/*
2 * Code related to writing an iolog of what a thread is doing, and to
3 * later read that back and replay
4 */
5#include <stdio.h>
6#include <stdlib.h>
7#include <libgen.h>
8#include <assert.h>
9#include "flist.h"
10#include "fio.h"
11#include "verify.h"
12
13static const char iolog_ver2[] = "fio version 2 iolog";
14
15void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
16{
17 flist_add_tail(&ipo->list, &td->io_log_list);
18 td->total_io_size += ipo->len;
19}
20
21void log_io_u(struct thread_data *td, struct io_u *io_u)
22{
23 const char *act[] = { "read", "write", "sync", "datasync",
24 "sync_file_range" };
25
26 assert(io_u->ddir <= 4);
27
28 if (!td->o.write_iolog_file)
29 return;
30
31 fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
32 act[io_u->ddir], io_u->offset,
33 io_u->buflen);
34}
35
36void log_file(struct thread_data *td, struct fio_file *f,
37 enum file_log_act what)
38{
39 const char *act[] = { "add", "open", "close" };
40
41 assert(what < 3);
42
43 if (!td->o.write_iolog_file)
44 return;
45
46
47 /*
48 * this happens on the pre-open/close done before the job starts
49 */
50 if (!td->iolog_f)
51 return;
52
53 fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
54}
55
56static void iolog_delay(struct thread_data *td, unsigned long delay)
57{
58 unsigned long usec = utime_since_now(&td->last_issue);
59
60 if (delay < usec)
61 return;
62
63 delay -= usec;
64
65 /*
66 * less than 100 usec delay, just regard it as noise
67 */
68 if (delay < 100)
69 return;
70
71 usec_sleep(td, delay);
72}
73
74static int ipo_special(struct thread_data *td, struct io_piece *ipo)
75{
76 struct fio_file *f;
77 int ret;
78
79 /*
80 * Not a special ipo
81 */
82 if (ipo->ddir != DDIR_INVAL)
83 return 0;
84
85 f = td->files[ipo->fileno];
86
87 switch (ipo->file_action) {
88 case FIO_LOG_OPEN_FILE:
89 ret = td_io_open_file(td, f);
90 if (!ret)
91 break;
92 td_verror(td, ret, "iolog open file");
93 return -1;
94 case FIO_LOG_CLOSE_FILE:
95 td_io_close_file(td, f);
96 break;
97 case FIO_LOG_UNLINK_FILE:
98 unlink(f->file_name);
99 break;
100 default:
101 log_err("fio: bad file action %d\n", ipo->file_action);
102 break;
103 }
104
105 return 1;
106}
107
108int read_iolog_get(struct thread_data *td, struct io_u *io_u)
109{
110 struct io_piece *ipo;
111 unsigned long elapsed;
112
113 while (!flist_empty(&td->io_log_list)) {
114 int ret;
115
116 ipo = flist_entry(td->io_log_list.next, struct io_piece, list);
117 flist_del(&ipo->list);
118
119 ret = ipo_special(td, ipo);
120 if (ret < 0) {
121 free(ipo);
122 break;
123 } else if (ret > 0) {
124 free(ipo);
125 continue;
126 }
127
128 io_u->ddir = ipo->ddir;
129 if (ipo->ddir != DDIR_WAIT) {
130 io_u->offset = ipo->offset;
131 io_u->buflen = ipo->len;
132 io_u->file = td->files[ipo->fileno];
133 get_file(io_u->file);
134
135 dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
136 io_u->buflen, io_u->file->file_name);
137 if (ipo->delay) iolog_delay(td, ipo->delay);
138 } else {
139 elapsed = mtime_since_genesis();
140 if (ipo->delay > elapsed)
141 usec_sleep(td, (ipo->delay - elapsed) * 1000);
142
143 }
144
145 free(ipo);
146
147 if (ipo->ddir != DDIR_WAIT)
148 return 0;
149 }
150
151 td->done = 1;
152 return 1;
153}
154
155void prune_io_piece_log(struct thread_data *td)
156{
157 struct io_piece *ipo;
158 struct rb_node *n;
159
160 while ((n = rb_first(&td->io_hist_tree)) != NULL) {
161 ipo = rb_entry(n, struct io_piece, rb_node);
162 rb_erase(n, &td->io_hist_tree);
163 free(ipo);
164 }
165
166 while (!flist_empty(&td->io_hist_list)) {
167 ipo = flist_entry(td->io_hist_list.next, struct io_piece, list);
168 flist_del(&ipo->list);
169 free(ipo);
170 }
171}
172
173/*
174 * log a successful write, so we can unwind the log for verify
175 */
176void log_io_piece(struct thread_data *td, struct io_u *io_u)
177{
178 struct rb_node **p, *parent;
179 struct io_piece *ipo, *__ipo;
180
181 ipo = malloc(sizeof(struct io_piece));
182 ipo->file = io_u->file;
183 ipo->offset = io_u->offset;
184 ipo->len = io_u->buflen;
185
186 /*
187 * We don't need to sort the entries, if:
188 *
189 * Sequential writes, or
190 * Random writes that lay out the file as it goes along
191 *
192 * For both these cases, just reading back data in the order we
193 * wrote it out is the fastest.
194 *
195 * One exception is if we don't have a random map AND we are doing
196 * verifies, in that case we need to check for duplicate blocks and
197 * drop the old one, which we rely on the rb insert/lookup for
198 * handling.
199 */
200 if ((!td_random(td) || !td->o.overwrite) &&
201 (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
202 INIT_FLIST_HEAD(&ipo->list);
203 flist_add_tail(&ipo->list, &td->io_hist_list);
204 return;
205 }
206
207 RB_CLEAR_NODE(&ipo->rb_node);
208
209 /*
210 * Sort the entry into the verification list
211 */
212restart:
213 p = &td->io_hist_tree.rb_node;
214 parent = NULL;
215 while (*p) {
216 parent = *p;
217
218 __ipo = rb_entry(parent, struct io_piece, rb_node);
219 if (ipo->offset < __ipo->offset)
220 p = &(*p)->rb_left;
221 else if (ipo->offset > __ipo->offset)
222 p = &(*p)->rb_right;
223 else {
224 assert(ipo->len == __ipo->len);
225 rb_erase(parent, &td->io_hist_tree);
226 goto restart;
227 }
228 }
229
230 rb_link_node(&ipo->rb_node, parent, p);
231 rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
232}
233
234void write_iolog_close(struct thread_data *td)
235{
236 fflush(td->iolog_f);
237 fclose(td->iolog_f);
238 free(td->iolog_buf);
239 td->iolog_f = NULL;
240 td->iolog_buf = NULL;
241}
242
243/*
244 * Read version 2 iolog data. It is enhanced to include per-file logging,
245 * syncs, etc.
246 */
247static int read_iolog2(struct thread_data *td, FILE *f)
248{
249 unsigned long long offset;
250 unsigned int bytes;
251 int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
252 char *fname, *act;
253 char *str, *p;
254 enum fio_ddir rw;
255
256 free_release_files(td);
257
258 /*
259 * Read in the read iolog and store it, reuse the infrastructure
260 * for doing verifications.
261 */
262 str = malloc(4096);
263 fname = malloc(256+16);
264 act = malloc(256+16);
265
266 reads = writes = waits = 0;
267 while ((p = fgets(str, 4096, f)) != NULL) {
268 struct io_piece *ipo;
269 int r;
270
271 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
272 &bytes);
273 if (r == 4) {
274 /*
275 * Check action first
276 */
277 if (!strcmp(act, "wait"))
278 rw = DDIR_WAIT;
279 else if (!strcmp(act, "read"))
280 rw = DDIR_READ;
281 else if (!strcmp(act, "write"))
282 rw = DDIR_WRITE;
283 else if (!strcmp(act, "sync"))
284 rw = DDIR_SYNC;
285 else if (!strcmp(act, "datasync"))
286 rw = DDIR_DATASYNC;
287 else {
288 log_err("fio: bad iolog file action: %s\n",
289 act);
290 continue;
291 }
292 } else if (r == 2) {
293 rw = DDIR_INVAL;
294 if (!strcmp(act, "add")) {
295 td->o.nr_files++;
296 fileno = add_file(td, fname);
297 file_action = FIO_LOG_ADD_FILE;
298 continue;
299 } else if (!strcmp(act, "open")) {
300 fileno = get_fileno(td, fname);
301 file_action = FIO_LOG_OPEN_FILE;
302 } else if (!strcmp(act, "close")) {
303 fileno = get_fileno(td, fname);
304 file_action = FIO_LOG_CLOSE_FILE;
305 } else {
306 log_err("fio: bad iolog file action: %s\n",
307 act);
308 continue;
309 }
310 } else {
311 log_err("bad iolog2: %s", p);
312 continue;
313 }
314
315 if (rw == DDIR_READ)
316 reads++;
317 else if (rw == DDIR_WRITE) {
318 /*
319 * Don't add a write for ro mode
320 */
321 if (read_only)
322 continue;
323 writes++;
324 } else if (rw == DDIR_WAIT) {
325 waits++;
326 } else if (rw == DDIR_INVAL) {
327 } else if (!ddir_sync(rw)) {
328 log_err("bad ddir: %d\n", rw);
329 continue;
330 }
331
332 /*
333 * Make note of file
334 */
335 ipo = malloc(sizeof(*ipo));
336 memset(ipo, 0, sizeof(*ipo));
337 INIT_FLIST_HEAD(&ipo->list);
338 ipo->ddir = rw;
339 if (rw == DDIR_WAIT) {
340 ipo->delay = offset;
341 } else {
342 ipo->offset = offset;
343 ipo->len = bytes;
344 if (bytes > td->o.max_bs[rw])
345 td->o.max_bs[rw] = bytes;
346 ipo->fileno = fileno;
347 ipo->file_action = file_action;
348 }
349
350 queue_io_piece(td, ipo);
351 }
352
353 free(str);
354 free(act);
355 free(fname);
356
357 if (writes && read_only) {
358 log_err("fio: <%s> skips replay of %d writes due to"
359 " read-only\n", td->o.name, writes);
360 writes = 0;
361 }
362
363 if (!reads && !writes && !waits)
364 return 1;
365 else if (reads && !writes)
366 td->o.td_ddir = TD_DDIR_READ;
367 else if (!reads && writes)
368 td->o.td_ddir = TD_DDIR_WRITE;
369 else
370 td->o.td_ddir = TD_DDIR_RW;
371
372 return 0;
373}
374
375/*
376 * open iolog, check version, and call appropriate parser
377 */
378static int init_iolog_read(struct thread_data *td)
379{
380 char buffer[256], *p;
381 FILE *f;
382 int ret;
383
384 f = fopen(td->o.read_iolog_file, "r");
385 if (!f) {
386 perror("fopen read iolog");
387 return 1;
388 }
389
390 p = fgets(buffer, sizeof(buffer), f);
391 if (!p) {
392 td_verror(td, errno, "iolog read");
393 log_err("fio: unable to read iolog\n");
394 return 1;
395 }
396
397 /*
398 * version 2 of the iolog stores a specific string as the
399 * first line, check for that
400 */
401 if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
402 ret = read_iolog2(td, f);
403 else {
404 log_err("fio: iolog version 1 is no longer supported\n");
405 ret = 1;
406 }
407
408 fclose(f);
409 return ret;
410}
411
412/*
413 * Setup a log for storing io patterns.
414 */
415static int init_iolog_write(struct thread_data *td)
416{
417 struct fio_file *ff;
418 FILE *f;
419 unsigned int i;
420
421 f = fopen(td->o.write_iolog_file, "a");
422 if (!f) {
423 perror("fopen write iolog");
424 return 1;
425 }
426
427 /*
428 * That's it for writing, setup a log buffer and we're done.
429 */
430 td->iolog_f = f;
431 td->iolog_buf = malloc(8192);
432 setvbuf(f, td->iolog_buf, _IOFBF, 8192);
433
434 /*
435 * write our version line
436 */
437 if (fprintf(f, "%s\n", iolog_ver2) < 0) {
438 perror("iolog init\n");
439 return 1;
440 }
441
442 /*
443 * add all known files
444 */
445 for_each_file(td, ff, i)
446 log_file(td, ff, FIO_LOG_ADD_FILE);
447
448 return 0;
449}
450
451int init_iolog(struct thread_data *td)
452{
453 int ret = 0;
454
455 if (td->o.read_iolog_file) {
456 /*
457 * Check if it's a blktrace file and load that if possible.
458 * Otherwise assume it's a normal log file and load that.
459 */
460 if (is_blktrace(td->o.read_iolog_file))
461 ret = load_blktrace(td, td->o.read_iolog_file);
462 else
463 ret = init_iolog_read(td);
464 } else if (td->o.write_iolog_file)
465 ret = init_iolog_write(td);
466
467 return ret;
468}
469
470void setup_log(struct io_log **log)
471{
472 struct io_log *l = malloc(sizeof(*l));
473
474 l->nr_samples = 0;
475 l->max_samples = 1024;
476 l->log = malloc(l->max_samples * sizeof(struct io_sample));
477 *log = l;
478}
479
480void __finish_log(struct io_log *log, const char *name)
481{
482 unsigned int i;
483 FILE *f;
484
485 f = fopen(name, "a");
486 if (!f) {
487 perror("fopen log");
488 return;
489 }
490
491 for (i = 0; i < log->nr_samples; i++) {
492 fprintf(f, "%lu, %lu, %u, %u\n", log->log[i].time,
493 log->log[i].val,
494 log->log[i].ddir,
495 log->log[i].bs);
496 }
497
498 fclose(f);
499 free(log->log);
500 free(log);
501}
502
503void finish_log_named(struct thread_data *td, struct io_log *log,
504 const char *prefix, const char *postfix)
505{
506 char file_name[256], *p;
507
508 snprintf(file_name, 200, "%s_%s.log", prefix, postfix);
509 p = basename(file_name);
510 __finish_log(log, p);
511}
512
513void finish_log(struct thread_data *td, struct io_log *log, const char *name)
514{
515 finish_log_named(td, log, td->o.name, name);
516}